Osprey
ROOSTSafety rules engine for real-time abuse mitigation, originally built at Discord and donated to ROOST. It evaluates events and their properties in real time using a Python-based rules DSL called SML, ingesting signals including machine-learning classifier outputs to detect and action spam, abuse and policy violations. It is Trust and Safety operational infrastructure rather than an AI guardrail model.
Openness
5 high confidence- license
- Apache-2.0(OSI)
- source
- public
- self-host
- yes
- service
- none
- core features withheld
- no
The rules engine, investigation console and plugin system are all published as open source, and adopters run the whole stack on their own infrastructure. ROOST, which maintains it, sells no hosted or enterprise edition, and ML models plug in as the adopter's own add-ons rather than as a gated component.
- https://docs.roost.tools/osprey/latest/ recorded 2026-09-29
The documentation welcome page describes "an open source rules engine and investigation console for trust and safety teams" that is "now developed in the open by" ROOST, and points readers to GitHub "for source code and project information". Searched the body for enterprise, license key, pricing and paid: none.
- https://docs.roost.tools/osprey/latest/development/ recorded 2026-09-29
The Getting Started guide brings up the whole stack locally with Docker Compose, run "from a clone of the repo", and lists "Kafka, Druid, PostgreSQL, MinIO, and" the project's own worker, API and UI as the containers it starts. Searched the body for license key, enterprise, hosted and pricing: none.
- https://docs.roost.tools/osprey/latest/integration/integrations.html recorded 2026-09-29
The Integrations and Plugins guide says Osprey is "designed to be extended without modifying the core codebase" and that ML models are optional add-ons: "ML models can be integrated as UDFs", with the example loading the adopter's own model file. No bundled model weights, license key or enterprise build are described. Searched the body for enterprise, license key and pricing: none.
- https://github.com/roostorg/osprey recorded 2026-08-13
Repo metadata records license spdxId - Apache-2.0. README brings the full stack up with one Docker Compose command (./demo.sh) and describes it as 'a working system, not a prototype', pointing at the development guide for building it yourself; ROOST is a non-profit and there is no paid tier, enterprise directory or license key on the page.
- https://roost.tools/blog/introducing-osprey-v1-0-open-source-infrastructure-for-real-time-abuse-mitigation/ recorded 2026-09-29
ROOST's V1.0 announcement, published January 28, 2026, states Osprey "is open source with an Apache 2 license" and that "any organization can use it for free on their own infrastructure, inspect how it works, adapt it to their needs, and contribute improvements back to the community". Searched the body for enterprise, pricing, paid and license key: none.
- https://roost.tools/osprey recorded 2026-09-29
ROOST's Osprey project page says "As an open source, self-hosted solution, data never leaves your environment and is never shared with third parties", and that Osprey "is developed and maintained by ROOST, a nonprofit building open source trust and safety infrastructure". Its only deployment call to action is a contact form. Searched the body for pricing, enterprise, paid and license key: none.
Adoption
3 medium confidenceOsprey's own README lists no adopters, so the repository is not where the adoption evidence sits. ROOST's launch announcement documents production use instead, naming Discord, Bluesky and the Matrix.org Foundation as deployments, though it gives no count of users or installs.
- https://github.com/roostorg/osprey recorded 2026-08-13
stargazerCount 462 for roostorg/osprey. The README's 'Adopters' section reads 'Osprey is used by:' with nothing under it.
- https://roost.tools/blog/introducing-osprey-v1-0-open-source-infrastructure-for-real-time-abuse-mitigation/ recorded 2026-10-08
V1.0 announcement, published January 28, 2026. Osprey 'processes over 50M events daily in production environments like Bluesky and is designed for real-time incident response'. Named production users with attributed quotes - Discord (Savannah Badalich, Global Head of Product Policy), Bluesky (Aaron Rodericks, Head of Trust and Safety) and The Matrix.org Foundation (Jim Mackenzie, VP of Trust & Safety). No download, install or customer count is published.
Capability
4 medium confidenceOsprey is a real-time rules engine and investigation console, built at Discord, where rules are written in SML and extended with user-defined functions. It is a general Trust and Safety engine rather than an AI-native guardrail, so its capability set reflects operational rule evaluation rather than model-specific safety detection.
- https://docs.roost.tools/osprey/latest/concepts.html recorded 2026-09-29
The Concepts page says Osprey "watches a stream of events happening on your platform, runs each one through your rules in real time, and records what it found so analysts can query and act on it". A rule "is a named condition over features, written in" SML, and effects include actions that "ban the user, add a label to an entity, declare a verdict".
- https://docs.roost.tools/osprey/latest/integration/integrations.html recorded 2026-09-29
The Integrations and Plugins guide says a UDF "is a Python class that can be called from your rules" and that UDFs "encapsulate reusable detection logic" including ML inference, while "ML models can be integrated as UDFs" rather than being built into the engine.
- https://docs.roost.tools/osprey/latest/rules/index.html recorded 2026-09-29
The Writing Rules page says "Osprey rules are written in SML, a subset of Python with additional restrictions to simplify rule writing", and that a UDF is "a function implemented in Python and made available to SML", with a shipped standard library and custom ones registered through plugins.
- https://github.com/roostorg/osprey recorded 2026-08-13
README describes 'a safety rules engine and investigation console for real-time event processing at scale' where 'human-written rules evaluate each one as it arrives', with rules 'written in SML, Osprey's structured rule language, and extended with user-defined functions (UDFs)', originally built at Discord. It carries no throughput figures and no longer describes the Rust/Python worker split.
- https://github.com/roostorg/osprey recorded 2026-08-14
The README carries no occurrence of "2.3M" or "400M" anywhere. It describes "a safety rules engine and investigation console for real-time event processing at scale", with rules written in SML and extended with user-defined functions.
- https://roost.tools/blog/introducing-osprey-v1-0-open-source-infrastructure-for-real-time-abuse-mitigation/ recorded 2026-09-29
ROOST's V1.0 announcement says Osprey "processes over 50M events daily in production environments like" Bluesky and "is designed for real-time incident response", and that the refactored open source version "has been pulled back into" Discord.
- https://roost.tools/osprey recorded 2026-09-29
ROOST's Osprey project page reports from its Bluesky case study that Osprey "now processes over 45 million events daily with 100,000+ automated enforcement actions", and describes it as "A high-performance investigation and rules engine for detecting and responding to real-time abuses at scale".
Verified 2026-09-29